iHarmony4

Dataset Overview

Dataset Name

iHarmony4

Dataset Description

iHarmony4 is the first large‑scale image harmonization dataset, aiming to adjust the appearance of synthetic image foregrounds to be consistent with background regions. The dataset includes four sub‑datasets: HCOCO, HAdobe5k, HFlickr, and Hday2night, each containing synthetic composite images, foreground masks, and corresponding real images.

Sub‑Dataset Details

1. HCOCO

• Source: Based on the Microsoft COCO dataset.
• Content: Contains 42 k synthetic composite images, edited using various color transfer methods on the foreground region.
• Storage: Available at Baidu Cloud (access code: ab5e) and OneDrive.

2. HAdobe5k

• Source: Based on the MIT‑Adobe FiveK dataset.
• Content: Foreground regions manually segmented and swapped between two versions.
• Storage: Available at Baidu Cloud and OneDrive.

3. HFlickr

• Source: Collected from Flickr (4,833 images).
• Content: Manually segmented foregrounds, generated using the same method as HCOCO.
• Storage: Available at Baidu Cloud and OneDrive.

4. Hday2night

• Source: Based on the day2night dataset.
• Content: Manually segmented foregrounds, cropped and overlaid onto another image captured at a different time.
• Storage: Available at Baidu Cloud and OneDrive.

Dataset Statistics

Dataset Storage

The iHarmony4 dataset is provided at Baidu Cloud (access code: kqz3) and OneDrive.